Automatic expert cohorts

The publisher controller now runs successive native training jobs and continues after rejected quality. The operated integration also covers publisher restart, rejected repeated source data and an accepted model that keeps answering while new cohorts execute.

The complete evidence release contains the source, numerical objects, original signed blocks, final state, reproduction drivers and the earlier failed driver attempt.

Measured operation	Result
Native consensus	Four CometBFT validators; one host and administrator
Automatically executed cohorts	Two, under one unchanged genesis
Numerical work	Real prefix production, two updates per cohort, three fresh audits per claim
Issuance	Exactly four trial NEURO for four verified updates
Quality	Both correctly measured failures rejected; serving root unchanged
Data rejection	Repeated original documents refused before a new proposal
Recovery	Publisher journal and signer reopened after the first rejection
Serving during work	Nine identical responses from persistent accepted shard processes
Validator agreement	Four matching app hashes at the final commitment height
Full ledger replay	All 74 transactions, 449 headers, exact final state
Operated duration	406.98 seconds, plus export/replay

These are small synthetic models and public test identities. The nine serving probes check continued execution, not useful language answers or paid inference. Correct training can earn the prescribed trial reward while a failed quality gate keeps those weights out of serving. The separate larger-model lifecycle evidence is recorded under task 3 in TODO.

Execution and recovery

neuroshard.evolution.expert_operator.Operator follows committed state and journals preparation, funding, reservations, numerical requests and signed transactions. The preparation and execution commands are operator-installed programs; network data reaches them as bounded JSON. Curators and auditors retain their own signing authority. Mechanical source review does not itself approve a cohort or establish that its content is trustworthy.

The operated trial used expert_preparation to read consecutive source windows, preserve tokenization and provenance, check prior history, and seal the next job. Actual prefix/training executors produced the claims. Separate numerical runs replayed those claims and the complete quality evaluation before audit verdicts. The rejected candidate closed normally, allowing preparation of the next cohort.

The first driver attempt waited for its original audit deadline after all bonded auditors had committed and consensus had shortened the window. It missed reveal and received no training reward. The corrected driver reads the current deadline. The audit daemon also refreshes state after replay. An unknown signed audit may be retired only when trusted committed history identifies its permanently closed context; its original bytes and unknown outcome remain journaled. CheckTx errors and a merely absent current claim cannot release its pending nonce.

Reproduction and remaining integration

The release binds source commit 130a2e2598b4b9ecbdb9a4d03663f51275ce2e00 and includes every numerical artifact for the recorded run. The integration test is tests/evolution/test_expert_operator_numerical.py. The archived operate.py driver runs that same numerical workflow against disposable CometBFT 0.38.26 validators on ports 39900–39932. It requires a fresh output directory, the pinned native/numerical dependencies and sufficient RAM for concurrent shard processes. Set NEUROSHARD_COHORT_SOURCE to the extracted source checkout and NEUROSHARD_COMETBFT to the pinned binary; then run the driver from a fresh directory. export_replay.py restarts those disposable validators, exports their public blocks and checks the complete application replay. Private validator keys are generated locally and are excluded from the published evidence.

Tasks 1 and 2 remain open. The source-selection and numerical commands in this integration use local synthetic fixtures. Connecting the controller to a live immutable feed and deployed GPU owners, and automatically promoting successive candidates that pass cumulative LLM answer-quality gates, remain required.

Structured immutable source feeds

expert_source.collect publishes bounded windows of original conversation messages, preserving roles, source revision, license and row positions. Held-out windows additionally preserve their scoring metadata. Each window and feed head is content-addressed; published updates append to the previous inventory. A crash after upload but before journal acknowledgement recreates the same objects and does not skip rows. Publication and native admission have separate cursors.

expert_source.Feed reads directly from the existing local or S3 object stores and is the upstream argument accepted by expert_preparation.prepare and seal. It verifies hashes, byte bounds, row counts and contiguous windows. A running reader accepts only an extension of its pinned head. Missing data raises an availability error rather than becoming a rejected or silently skipped row. The curator still checks allowed publishers, original source correspondence, tokenization, historical duplicates and contamination before approving admission. An immutable feed is not evidence that its publisher's content is true or safe.

Five checks cover cross-window reads, publisher and reader restart, an upload-before-journal crash, rollback/overlap rejection, corrupted or missing objects, bounded S3 reads and actual native proposal preparation and review. The semantic trial's 192 training conversations were also published in three 64-row windows: nine immutable S3 objects, 42,090 bytes, all verified through full public downloads. The reconstructed conversations exactly match the training source; this publication opens no evaluation set and admits no native job.

A deployment's installed preparation command can consume this reader and return the sealed job to the existing Operator. The complete integration with deployed GPU owners and passing cumulative LLM quality remains outstanding; the collector and feed reader do not claim that integration by themselves.